Prediction of floods using improved PCA with one-dimensional convolutional neural network

Abstract

Forecasting floods have always been a difficult task due to the complexity of the available data. Machine learning techniques have been widely used to predict floods based on precipitation, humidity, temperature, water velocity, and level variables. However, most prior studies have examined the monthly rainfall intensity to determine the likelihood of flooding. As a result, a state's daily and monthly rainfall intensity has been used to train deep-learning models to predict floods. In addition, feature reduction approaches are critical for dealing with data of a large dimensionality and improving classification accuracy. This article utilizes improved Principal Component Analysis (i-PCA), a linear unsupervised statistical transformation, as a feature reduction procedure. A 1D-Convolutional Neural Network (CNN) model forecasts the flood based on the reduced features. The experiments are based on a dataset of daily and monthly rainfall data collected from 1901 to 2021 for Kerala state. Qualitative analysis is performed using precision, accuracy, recall and F1-score parameters. The experiment analysis proves that the proposed algorithm attained 94.24% accuracy, and existing techniques achieved 86% of accuracy performance. The reason is that the proposed model uses the improved PCA for the feature reduction technique.